HealthEdAI: An End-to-End Agentic Health Education Platform - Combining LLaMA-3 Inference, Retrieval-Augmented Generation, and Multi-Role Clinical Workflows

The issue of medical expertise versus patient comprehension persists as an ongoing challenge in the discipline of medicine. In many places, patients are provided with medical information concerning their ailments that might not be easily comprehensible to them. There are clear consequences to this situation, including instances where patients fail to take their prescribed medications, medical delays, and avoidable deterioration of medical conditions. The current digital health interventions that exist in our environment have not been successful in solving this challenge. This could be due to the supply of generic medical information that does not take into consideration different health literacies among patients. The current paper introduces HealthEdAI, which is a production-ready multi-task patient education framework based on the integration of three closely related AI sub-systems: inferencing of the pretrained LLaMA-3.1-8B model using Groq free tier API, crawler for the RAG system, and a ReAct agent empowered by LangChain and utilizing three auxiliary utilities. The relevance checking is one of the key innovations introduced in this study. Rather than blindly returning any retrieved document from the RAG system to the language model, the ReAct agent performs a preliminary verification if it covers the specified disease at all; otherwise, parametric knowledge is used. Thanks to the above, response generation quality was boosted up to 87% for rare diseases compared to 12%. Alongside the question-answering facility, the system offers many clinically-oriented services for urgent symptom assessment, drug interaction detection, laboratory reports, physical activity indicators, appointment scheduling, reminders regarding medication dosage, multilingual response translation, and immediate location-based identification of nearest medical professionals through Practo. In total, there are three different user types that are pre-defined upon registration: patient, doctor, and administrator. The evaluation metrics calculated for 120 health queries include FleschKincaid Grade Level of 6.2, ROUGE-1 F1 of 0.74, BLEU-4 of 0.61, and 100% safety rating, outperforming all three baseline models.